ARM: teach AAPCS-VFP to deal with Cortex-M4.

Cortex-M4 only has single-precision floating point support, so any LLVM
"double" type will have been split into 2 i32s by now. Fortunately, the
consecutive-register framework turns out to be precisely what's needed to
reconstruct the double and follow AAPCS-VFP correctly!

rdar://problem/17012966

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@209650 91177308-0d34-0410-b5e6-96231b3b80d8

4 files changed, 143 insertions, 21 deletions

diff --git a/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp b/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp
index c181046ba23..070e929fce7 100644
--- a/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp
+++ b/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp
@@ -7176,11 +7176,8 @@ TargetLowering::LowerCallTo(TargetLowering::CallLoweringInfo &CLI) const {
       }
       if (Args[i].isNest)
         Flags.setNest();
-      if (NeedsRegBlock) {
+      if (NeedsRegBlock)
         Flags.setInConsecutiveRegs();
-        if (Value == NumValues - 1)
-          Flags.setInConsecutiveRegsLast();
-      }
       Flags.setOrigAlign(OriginalAlignment);
 
       MVT PartVT = getRegisterType(CLI.RetTy->getContext(), VT);
@@ -7226,6 +7223,10 @@ TargetLowering::LowerCallTo(TargetLowering::CallLoweringInfo &CLI) const {
         else if (j != 0)
           MyFlags.Flags.setOrigAlign(1);
 
+        // Only mark the end at the last register of the last value.
+        if (NeedsRegBlock && Value == NumValues - 1 && j == NumParts - 1)
+          MyFlags.Flags.setInConsecutiveRegsLast();
+
         CLI.Outs.push_back(MyFlags);
         CLI.OutVals.push_back(Parts[j]);
       }
@@ -7412,11 +7413,8 @@ void SelectionDAGISel::LowerArguments(const Function &F) {
       }
       if (F.getAttributes().hasAttribute(Idx, Attribute::Nest))
         Flags.setNest();
-      if (NeedsRegBlock) {
+      if (NeedsRegBlock)
         Flags.setInConsecutiveRegs();
-        if (Value == NumValues - 1)
-          Flags.setInConsecutiveRegsLast();
-      }
       Flags.setOrigAlign(OriginalAlignment);
 
       MVT RegisterVT = TLI->getRegisterType(*CurDAG->getContext(), VT);
@@ -7429,6 +7427,11 @@ void SelectionDAGISel::LowerArguments(const Function &F) {
         // if it isn't first piece, alignment must be 1
         else if (i > 0)
           MyFlags.Flags.setOrigAlign(1);
+
+        // Only mark the end at the last register of the last value.
+        if (NeedsRegBlock && Value == NumValues - 1 && i == NumRegs - 1)
+          MyFlags.Flags.setInConsecutiveRegsLast();
+
         Ins.push_back(MyFlags);
       }
       PartBase += VT.getStoreSize();
diff --git a/lib/Target/ARM/ARMCallingConv.h b/lib/Target/ARM/ARMCallingConv.h
index 8e0fd893528..dc41c1c14bb 100644
--- a/lib/Target/ARM/ARMCallingConv.h
+++ b/lib/Target/ARM/ARMCallingConv.h
@@ -177,9 +177,8 @@ static bool CC_ARM_AAPCS_Custom_HA(unsigned &ValNo, MVT &ValVT, MVT &LocVT,
                                    CCValAssign::LocInfo &LocInfo,
                                    ISD::ArgFlagsTy &ArgFlags, CCState &State) {
   SmallVectorImpl<CCValAssign> &PendingHAMembers = State.getPendingLocs();
-
   // AAPCS HFAs must have 1-4 elements, all of the same type
-  assert(PendingHAMembers.size() < 4);
+  assert(PendingHAMembers.size() < 8);
   if (PendingHAMembers.size() > 0)
     assert(PendingHAMembers[0].getLocVT() == LocVT);
 
@@ -189,7 +188,7 @@ static bool CC_ARM_AAPCS_Custom_HA(unsigned &ValNo, MVT &ValVT, MVT &LocVT,
       CCValAssign::getPending(ValNo, ValVT, LocVT, LocInfo));
 
   if (ArgFlags.isInConsecutiveRegsLast()) {
-    assert(PendingHAMembers.size() > 0 && PendingHAMembers.size() <= 4 &&
+    assert(PendingHAMembers.size() > 0 && PendingHAMembers.size() <= 8 &&
            "Homogeneous aggregates must have between 1 and 4 members");
 
     // Try to allocate a contiguous block of registers, each of the correct
@@ -197,6 +196,7 @@ static bool CC_ARM_AAPCS_Custom_HA(unsigned &ValNo, MVT &ValVT, MVT &LocVT,
     const uint16_t *RegList;
     unsigned NumRegs;
     switch (LocVT.SimpleTy) {
+    case MVT::i32:
     case MVT::f32:
       RegList = SRegList;
       NumRegs = 16;
@@ -235,11 +235,20 @@ static bool CC_ARM_AAPCS_Custom_HA(unsigned &ValNo, MVT &ValVT, MVT &LocVT,
       State.AllocateReg(SRegList[regNo]);
 
     unsigned Size = LocVT.getSizeInBits() / 8;
-    unsigned Align = LocVT.SimpleTy == MVT::v2f64 ? 8 : Size;
+    unsigned Align = Size;
+
+    if (LocVT.SimpleTy == MVT::v2f64 || LocVT.SimpleTy == MVT::i32) {
+      // Vectors are always aligned to 8 bytes. If we've seen an i32 here
+      // it's because it's been split from a larger type, also with align 8.
+      Align = 8;
+    }
 
     for (auto It : PendingHAMembers) {
       It.convertToMem(State.AllocateStack(Size, Align));
       State.addLoc(It);
+
+      // Only the first member needs to be aligned.
+      Align = 1;
     }
 
     // All pending members have now been allocated
diff --git a/lib/Target/ARM/ARMISelLowering.cpp b/lib/Target/ARM/ARMISelLowering.cpp
index 5beb752d3a4..00d07e84067 100644
--- a/lib/Target/ARM/ARMISelLowering.cpp
+++ b/lib/Target/ARM/ARMISelLowering.cpp
@@ -10778,14 +10778,13 @@ static bool isHomogeneousAggregate(Type *Ty, HABaseType &Base,
 /// \brief Return true if a type is an AAPCS-VFP homogeneous aggregate.
 bool ARMTargetLowering::functionArgumentNeedsConsecutiveRegisters(
     Type *Ty, CallingConv::ID CallConv, bool isVarArg) const {
-  if (getEffectiveCallingConv(CallConv, isVarArg) ==
-      CallingConv::ARM_AAPCS_VFP) {
-    HABaseType Base = HA_UNKNOWN;
-    uint64_t Members = 0;
-    bool result = isHomogeneousAggregate(Ty, Base, Members);
-    DEBUG(dbgs() << "isHA: " << result << " "; Ty->dump(); dbgs() << "\n");
-    return result;
-  } else {
+  if (getEffectiveCallingConv(CallConv, isVarArg) !=
+      CallingConv::ARM_AAPCS_VFP)
     return false;
-  }
+
+  HABaseType Base = HA_UNKNOWN;
+  uint64_t Members = 0;
+  bool result = isHomogeneousAggregate(Ty, Base, Members);
+  DEBUG(dbgs() << "isHA: " << result << " "; Ty->dump(); dbgs() << "\n");
+  return result;
 }
diff --git a/test/CodeGen/ARM/aapcs-hfa-code.ll b/test/CodeGen/ARM/aapcs-hfa-code.ll
new file mode 100644
index 00000000000..396e83816cc
--- /dev/null
+++ b/test/CodeGen/ARM/aapcs-hfa-code.ll
@@ -0,0 +1,111 @@
+; RUN: llc < %s -mtriple=armv7-linux-gnueabihf -o - | FileCheck %s
+; RUN: llc < %s -mtriple=thumbv7em-none-eabi -mcpu=cortex-m4 | FileCheck %s --check-prefix=CHECK-M4F
+
+target datalayout = "e-m:e-p:32:32-i64:64-v128:64:128-n32-S64"
+
+define arm_aapcs_vfpcc void @test_1float({ float } %a) {
+  call arm_aapcs_vfpcc void @test_1float({ float } { float 1.0 })
+  ret void
+
+; CHECK-LABEL: test_1float:
+; CHECK-DAG: vmov.f32 s0, #1.{{0+}}e+00
+; CHECK: bl test_1float
+
+; CHECK-M4F-LABEL: test_1float:
+; CHECK-M4F-DAG: vmov.f32 s0, #1.{{0+}}e+00
+; CHECK-M4F: bl test_1float
+}
+
+define arm_aapcs_vfpcc void @test_2float({ float, float } %a) {
+  call arm_aapcs_vfpcc void @test_2float({ float, float } { float 1.0, float 2.0 })
+  ret void
+
+; CHECK-LABEL: test_2float:
+; CHECK-DAG: vmov.f32 s0, #1.{{0+}}e+00
+; CHECK-DAG: vmov.f32 s1, #2.{{0+}}e+00
+; CHECK: bl test_2float
+
+; CHECK-M4F-LABEL: test_2float:
+; CHECK-M4F-DAG: vmov.f32 s0, #1.{{0+}}e+00
+; CHECK-M4F-DAG: vmov.f32 s1, #2.{{0+}}e+00
+; CHECK-M4F: bl test_2float
+}
+
+define arm_aapcs_vfpcc void @test_3float({ float, float, float } %a) {
+  call arm_aapcs_vfpcc void @test_3float({ float, float, float } { float 1.0, float 2.0, float 3.0 })
+  ret void
+
+; CHECK-LABEL: test_3float:
+; CHECK-DAG: vmov.f32 s0, #1.{{0+}}e+00
+; CHECK-DAG: vmov.f32 s1, #2.{{0+}}e+00
+; CHECK-DAG: vmov.f32 s2, #3.{{0+}}e+00
+; CHECK: bl test_3float
+
+; CHECK-M4F-LABEL: test_3float:
+; CHECK-M4F-DAG: vmov.f32 s0, #1.{{0+}}e+00
+; CHECK-M4F-DAG: vmov.f32 s1, #2.{{0+}}e+00
+; CHECK-M4F-DAG: vmov.f32 s2, #3.{{0+}}e+00
+; CHECK-M4F: bl test_3float
+}
+
+define arm_aapcs_vfpcc void @test_1double({ double } %a) {
+; CHECK-LABEL: test_1double:
+; CHECK-DAG: vmov.f64 d0, #1.{{0+}}e+00
+; CHECK: bl test_1double
+
+; CHECK-M4F-LABEL: test_1double:
+; CHECK-M4F: movs [[ONEHI:r[0-9]+]], #0
+; CHECK-M4F: movs [[ONELO:r[0-9]+]], #0
+; CHECK-M4F: movt [[ONEHI]], #16368
+; CHECK-M4F-DAG: vmov s0, [[ONELO]]
+; CHECK-M4F-DAG: vmov s1, [[ONEHI]]
+; CHECK-M4F: bl test_1double
+
+  call arm_aapcs_vfpcc void @test_1double({ double } { double 1.0 })
+  ret void
+}
+
+; Final double argument might be put in s15 & [sp] if we're careless. It should
+; go all on the stack.
+define arm_aapcs_vfpcc void @test_1double_nosplit([4 x float], [4 x double], [3 x float], double %a) {
+; CHECK-LABEL: test_1double_nosplit:
+; CHECK-DAG: mov [[ONELO:r[0-9]+]], #0
+; CHECK-DAG: movw [[ONEHI:r[0-9]+]], #0
+; CHECK-DAG: movt [[ONEHI]], #16368
+; CHECK: strd [[ONELO]], [[ONEHI]], [sp]
+; CHECK: bl test_1double_nosplit
+
+; CHECK-M4F-LABEL: test_1double_nosplit:
+; CHECK-M4F: movs [[ONELO:r[0-9]+]], #0
+; CHECK-M4F: movs [[ONEHI:r[0-9]+]], #0
+; CHECK-M4F: movt [[ONEHI]], #16368
+; CHECK-M4F-DAG: str [[ONELO]], [sp]
+; CHECK-M4F-DAG: str [[ONEHI]], [sp, #4]
+; CHECK-M4F: bl test_1double_nosplit
+  call arm_aapcs_vfpcc void @test_1double_nosplit([4 x float] undef, [4 x double] undef, [3 x float] undef, double 1.0)
+  ret void
+}
+
+; Final double argument might go at [sp, #4] if we're careless. Should go at
+; [sp, #8] to preserve alignment.
+define arm_aapcs_vfpcc void @test_1double_misaligned([4 x double], [4 x double], float, double) {
+  call arm_aapcs_vfpcc void @test_1double_misaligned([4 x double] undef, [4 x double] undef, float undef, double 1.0)
+
+; CHECK-LABEL: test_1double_misaligned:
+; CHECK-DAG: mov [[ONELO:r[0-9]+]], #0
+; CHECK-DAG: mov r[[BASE:[0-9]+]], sp
+; CHECK-DAG: movw [[ONEHI:r[0-9]+]], #0
+; CHECK-DAG: movt [[ONEHI]], #16368
+; CHECK-DAG: str [[ONELO]], [r[[BASE]], #8]!
+; CHECK-DAG: str [[ONEHI]], [r[[BASE]], #4]
+
+; CHECK-M4F-LABEL: test_1double_misaligned:
+; CHECK-M4F: movs [[ONELO:r[0-9]+]], #0
+; CHECK-M4F: movs [[ONEHI:r[0-9]+]], #0
+; CHECK-M4F: movt [[ONEHI]], #16368
+; CHECK-M4F-DAG: str [[ONELO]], [sp, #8]
+; CHECK-M4F-DAG: str [[ONEHI]], [sp, #12]
+; CHECK-M4F: bl test_1double_misaligned
+
+  ret void
+}